| Re: 2D FFT Slow. Any ideas? [message #37234] |
Mon, 08 December 2003 00:02  |
Brian
Messages: 27
Registered: March 2001
|
Junior Member |
|
|
Hi,
The IDL code is
A=dcomplex(randomn(seed,2048, 2048,/DOUBLE), randomn(seed,2048,
2048,/DOUBLE))
t=systime(1)
FA=fft(A,-1)
print, systime(1)-t
The corresponding matlab code is
A = randn(2048,2048) + i*randn(2048,2048);
tic; FA=fft2(A); toc
I get around 10.4 sec in IDL, and around 3.7 sec in MATLAB 6.5
-brian
"R.G. Stockwell" <noemail@please.com> schrieb im Newsbeitrag
news:gN3Ab.767$o_.32178@news.uswest.net...
> "Brian" <brian.huether@NOdlrSPAM.de> wrote in message
> news:a298a85e9af4e70d51199dcae50c4c81@news.teranews.com...
>> I did a little benchmark between IDL and MATLAB. In each case I created
a
>> random double precision complex array of size 2048 by 2048 and timed how
>> long the 2D FFT took. In MATLAB 6.5 it took about 3.5 sec, and in IDL,
it
>> took about 10 sec. Is there a way to have IDL use MATLAB for the FFT,
>> perhaps using activex? Or would the overhead in using activex defeat the
>> purpose?
>>
>> thanks,
>>
>> brian
>>
>>
>
> You could try calling an external routine, some of the best being
available
> at (fastest ft in the west)
>
> http://fftw.org/
>
> Cheers,
> bob
>
> PS I can't believe active x calls would improve speed, but hey, you never
> know.
> And, I'm surprised that the canned IDL is not very fast. Any chance you
> don't really
> have a 2048^2 array in idl (did you make a 2049^2 array for instance?)
>
> If you post a blip of example code, I can run them here (matlab and idl)
and
> verify the time difference,
> which would be very interesting.
>
>
|
|
|
|
|
|
|
|
| Re: 2D FFT Slow. Any ideas? [message #37250 is a reply to message #37249] |
Fri, 05 December 2003 10:04  |
R.G. Stockwell
Messages: 363
Registered: July 1999
|
Senior Member |
|
|
"Brian" <brian.huether@NOdlrSPAM.de> wrote in message
news:a298a85e9af4e70d51199dcae50c4c81@news.teranews.com...
> I did a little benchmark between IDL and MATLAB. In each case I created a
> random double precision complex array of size 2048 by 2048 and timed how
> long the 2D FFT took. In MATLAB 6.5 it took about 3.5 sec, and in IDL, it
> took about 10 sec. Is there a way to have IDL use MATLAB for the FFT,
> perhaps using activex? Or would the overhead in using activex defeat the
> purpose?
>
> thanks,
>
> brian
>
>
You could try calling an external routine, some of the best being available
at (fastest ft in the west)
http://fftw.org/
Cheers,
bob
PS I can't believe active x calls would improve speed, but hey, you never
know.
And, I'm surprised that the canned IDL is not very fast. Any chance you
don't really
have a 2048^2 array in idl (did you make a 2049^2 array for instance?)
If you post a blip of example code, I can run them here (matlab and idl) and
verify the time difference,
which would be very interesting.
|
|
|
|
| Re: 2D FFT Slow. Any ideas? [message #37367 is a reply to message #37234] |
Mon, 08 December 2003 12:17 |
Richard French
Messages: 173
Registered: December 2000
|
Senior Member |
|
|
On 12/8/03 3:02 AM, in article
64b518634e4a83f7473363b5cd15989d@news.teranews.com, "Brian"
<brian.huether@NOdlrSPAM.de> wrote:
> Hi,
>
> The IDL code is
>
> A=dcomplex(randomn(seed,2048, 2048,/DOUBLE), randomn(seed,2048,
> 2048,/DOUBLE))
> t=systime(1)
> FA=fft(A,-1)
> print, systime(1)-t
>
> The corresponding matlab code is
>
> A = randn(2048,2048) + i*randn(2048,2048);
> tic; FA=fft2(A); toc
>
> I get around 10.4 sec in IDL, and around 3.7 sec in MATLAB 6.5
>
> -brian
>
I got FFTW working on a SUNBLADE 1000
IDL> print,!VERSION
{ sparc sunos unix Solaris 6.0 Jun 27 2003 32 64}
(I had to run IDL in 32-bit mode)
For a REAL FFT, 1024x1024, IDL took 2.9 times as long as FFTW
For 2048 x 2048, IDL took 3.4 times as long as FFTW.
On a Dec/HP/Compaq Alpha, the speedup was a factor of 4 for 1024x1024 and
about a factor of 5 for 2048x2048.
If you have lots of FFT's to compute, it is worth taking the trouble to
build a DLM for the FFTW routines. However, I did not find this easy to do,
I used the help from Stein Haugan, and I had to fiddle with the Makefile to
get everything working.
I am going to see if I can get fftw to work under MAC OS X, and to build a
DLM for that. If anyone has already succeeded in doing this, I'd be
interested in knowing how you did it.
Dick French
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
